The Fifth Promise

Stage 1 — The Conversational Interface

At the first level, AI changes how people interact with planning software. Instead of navigating through menus, reports and exception screens, a planner can simply ask: “Why are we going short on Product X next month?” The system can find the relevant information and explain what is happening.

This is already appearing across the market. SAP’s Supply Chain Planning Assistant, for example, is designed to help planners detect exceptions, investigate root causes, assess capacity and inventory, and simulate responses from a single workspace. Oracle is embedding AI agents into planning, procurement, manufacturing, inventory and logistics workflows.

This is unquestionably useful. Anyone who has worked with traditional planning software knows how much planner time is consumed finding information rather than making decisions. AI can dramatically improve that experience.

But there is an important distinction. At this stage, the underlying planning answer has not necessarily improved at all. We have simply made the existing answer easier to access and understand. That is valuable. But it is not autonomous planning.

Stage 2 — The Analytical Agent

The second stage moves beyond answering questions. The agent begins investigating them. A planner asks: “Why are we short?” The agent traces the issue through demand, inventory, purchase orders, production capacity, supplier lead times and other relevant data before identifying the likely cause.

This is where the productivity opportunity becomes much more interesting. A planner who previously spent two hours navigating multiple applications could potentially receive the analysis in seconds.

o9’s architecture is explicitly designed around connected context. Its Digital Brain is built around an Enterprise Knowledge Graph that connects data, models and decisions across strategic, tactical and operational horizons. Its APEX model then adds neuro-symbolic, generative and agentic AI capabilities on top of that foundation.

The concept is powerful. But there is a critical dependency: the agent can only reason from the reality represented in the system.

If a supplier lead time is wrong, machine capacity is stale, a production routing is incomplete, inventory is inaccurate or a maintenance outage is invisible, the agent does not somehow discover the physical truth. It reasons very efficiently from the information it has been given.

That distinction has haunted advanced planning systems for decades.

Stage 3 — The Scenario Agent

At the third stage, the agent stops merely explaining the current plan and starts evaluating alternatives. A planner might ask: “Can we move production from Plant A to Plant B?” The agent can invoke the planning engine, run scenarios and compare the consequences for service, inventory, capacity and cost.

OMP provides a particularly clear example of this architecture. Its UnisonIQ agents work from the customer’s digital twin and call the optimisation engines within Unison Planning. OMP explicitly states that the generative AI itself does not calculate the plan or decide independently; it orchestrates the underlying planning and optimisation capabilities.

That distinction is extremely important. The agent understands the request. The planning engine calculates the answer. The agent then interprets and presents the result.

This could make sophisticated scenario analysis dramatically easier to use. But it also exposes the historical problem again. If the underlying model contains thousands of assumptions about capacity, lead times, yields, routings, calendars, inventory, supplier performance and demand, how many of those assumptions accurately describe operational reality today?

The AI may make it much easier to run ten scenarios. That does not necessarily mean any of those ten scenarios are correct.

Stage 4 — Governed Action

This is where agentic planning begins to become genuinely different. Instead of merely recommending an action, the system can execute it within predefined authority. For example: “Reallocate this stock.” “Change this replenishment parameter.” “Release this order.” “Move this production quantity.”

Oracle is pushing strongly in this direction. Its Fusion Agentic Applications are designed to work inside existing enterprise permissions, policies, approval hierarchies and transactional context, allowing agents to progress work and execute decisions within defined guardrails. Its newer Fusion Claw architecture goes further, combining AI reasoning with deterministic enterprise computation to support simulation, modelling, continuous replanning and governed execution.

SAP is following a similar direction, with agents operating inside the existing business context and, when authorised, triggering coordinated action across planning, manufacturing, logistics, assets and supplier networks.

This is where the potential productivity improvement becomes substantial. A great deal of supply chain work consists of identifying something that is wrong, investigating it, determining the response, seeking approval, and then manually navigating multiple systems to implement it. Agents could remove a significant amount of that friction.

But the risk also changes. A poor recommendation is one thing. A poor recommendation that can execute itself is something else entirely.

That is why governance, permissions, auditability, explainability and human intervention are becoming increasingly important parts of serious agentic architectures.

Stage 5 — Autonomous Planning

The fifth stage is the vision generating much of the excitement: the system detects a problem, understands its consequences, generates alternatives, selects the best course of action and executes it — continuously and with minimal human intervention. This is the latest version of the autonomous supply chain.

And this is where history should make us cautious.

Gartner has warned specifically about “agent washing” in supply chain planning. Its view is that much of today’s agentic functionality improves conversational support, recommendations and user experience rather than fundamentally changing planning decisions. Gartner defines true autonomous planning much more demandingly: automatic generation of plans, automatic selection of the appropriate plan and seamless execution without human intervention. Most current solutions have not yet reached that level.

That is a useful reality check. Because there is a danger that the industry begins confusing a significantly better interface with a significantly better planning model.

We Have Been Here Before

Advanced Planning Systems have always had an appealing proposition: model the supply chain accurately, represent all the constraints, optimise the network, generate the best answer. The difficulty has rarely been the mathematics. The difficulty has been representing a constantly changing physical supply chain accurately enough for those mathematics to remain trustworthy.

Lead times change. Capacity is not what the master data says. Yield varies. Suppliers behave differently from contractual assumptions. Production schedules change. People override plans. Customers behave unexpectedly. Temporary constraints become permanent. Local workarounds accumulate. The theoretical model and the operational supply chain gradually diverge.

Organisations then add additional parameters, rules and constraints in an attempt to make the model more realistic. Eventually, the planning environment can become extremely sophisticated — and extremely difficult to maintain.

This is not a new problem. It is one of the reasons the history of APS contains so many projects that promised transformational planning but ultimately disappointed the people expected to use them. None of that disappears because the planner can now talk to an AI agent. In some respects, AI could make the problem more difficult to see.

AI Can Make a Bad Answer Sound Better

An old planning system might produce something like: Constraint violation – Resource 1847 – Period 42. Nobody is particularly seduced by that.

An AI agent might instead say: “I have identified a capacity constraint at Plant B next week. The optimal response is to transfer 12,000 units to Plant C, protecting customer service while minimising incremental logistics cost.” That sounds authoritative. It may explain its reasoning beautifully. It may present the answer in an impressive executive dashboard.

But suppose Plant C’s practical capacity is wrong. Suppose a changeover constraint has not been maintained. Suppose the required labour is unavailable. Suppose the relevant material has not passed quality control. Suppose a maintenance shutdown exists only in somebody’s local spreadsheet.

The fluency of the explanation has no bearing on the quality of the underlying model.

That may be one of the most important risks as the industry adopts agentic planning. AI does not necessarily fix a weak planning model. It may simply make a weak planning model sound considerably more authoritative.

That is why perhaps the most important question when evaluating agentic planning is not “How intelligent does the agent appear?” It is: “What data, assumptions and planning logic produced the answer underneath the agent?”

The Opportunity Is Still Very Real

None of this means agentic AI is a gimmick. Quite the opposite. There are potentially enormous gains from using agents to remove the administrative and analytical burden surrounding planning.

Imagine an agent that can identify an exception, investigate its causes, collect the relevant information, invoke the appropriate planning engine, run several scenarios, compare the outcomes, explain the trade-offs, route the recommendation for approval, and then implement an authorised response. That could remove hours of non-value-added planner activity.

OMP’s current architecture illustrates this particularly clearly. Its agents use established optimisation and planning engines rather than asking generative AI to invent a plan, while governance determines what an agent may recommend, execute or escalate. That feels like a much more credible future than simply asking a large language model to “optimise the supply chain”.

The technology can make planning faster, more accessible and significantly easier to navigate. But those benefits should not be confused with solving the fundamental modelling challenges that have existed for decades.

History Is Repeating — But It Doesn’t Have To

There is an interesting pattern in supply chain technology. Each generation identifies the shortcomings of the previous generation and introduces a new technology that promises to overcome them. MRP became MRP II. ERP became APS. APS evolved into concurrent planning. Then came digital twins, machine learning and autonomous planning. Now we have agentic AI.

Each development has contributed something valuable. But each has also tended to inherit many of the problems underneath it.

The danger is therefore not agentic AI itself. The danger is believing that attaching AI to an existing planning architecture somehow removes the architectural limitations that were already there. It doesn’t.

A system that required unrealistic levels of data accuracy yesterday does not suddenly require less accurate data because an agent sits on top of it today. A model that cannot represent operational reality does not become more representative because its output is delivered conversationally. And an optimisation engine generating an impractical answer does not become practical because an AI explains the answer elegantly.

If the industry forgets that, history may repeat itself once again.

The Fifth Promise

Every generation of planning technology has promised to overcome the limitations of the generation before it. Agentic AI may genuinely be one of the most important usability and orchestration breakthroughs supply chain planning has seen for years. But it does not repeal the laws of planning.

Bad data remains bad data. An unrealistic constraint remains unrealistic. An incomplete model remains incomplete. A mathematically optimal answer can still be operationally impossible.

The organisations that benefit most from agentic planning may therefore not be those with the most impressive AI interface. They may be those that understand where AI should stop — and where deterministic planning logic, optimisation, operational knowledge and human judgement must begin.

Perhaps the future of planning is not fully autonomous planning after all. Perhaps it is something more pragmatic: AI to understand the question. Planning engines to calculate the answer. Humans and governance to decide when the answer deserves to be trusted.

That may sound less revolutionary. History suggests it may also be considerably more useful.

How We Think About This at O8

At O8, we see agentic AI as an orchestration layer rather than a replacement for planning logic. Our approach is to keep underlying planning decisions grounded in established planning, simulation and execution engines, while using AI where it adds genuine value — understanding context, identifying patterns, investigating exceptions, navigating data and helping planners evaluate alternatives more quickly.

O8 Enterprise provides the operational planning and execution layer. It supports multiple planning approaches, including DDMRP, make-to-order, rate-based and other policy-driven methods, rather than assuming that one planning methodology is appropriate for every part of the supply chain.

Organic Planner provides network-wide planning, simulation and scenario analysis. It allows organisations to explore disruption, capacity, inventory and network decisions while using the same underlying O8 planning environment that supports operational execution.

Forecast Intelligence brings machine learning and probabilistic intelligence into demand forecasting and changing demand behaviour.

The natural role for agentic AI is to sit across those capabilities. An agent might understand the question, investigate the relevant data, invoke the appropriate planning or simulation process, compare alternatives and explain the result. But the planning engine should still produce the plan.

That distinction matters to us. We do not believe that a fluent AI response should ever be mistaken for a reliable planning answer. The quality of that answer still depends on the quality of the data, the assumptions, the underlying planning logic and the operational reality represented underneath it.

Our objective is therefore not to replace planners with AI. It is to remove as much of the non-value-added work around planning as possible — finding information, investigating exceptions, navigating systems and assembling alternatives — while keeping planning decisions explainable, governed and grounded in the underlying supply chain.

The future may well be agentic. But the intelligence of the agent will ultimately be judged by the quality of the planning environment beneath it.